TabPFN with zero training beats tuned XGBoost on 14 out of 14 tables
What happened
作者用 Grinsztajn 基准里的 14 个数据集，让 TabPFN 和 TabICL 跟调过参的 XGBoost 比了一场。这两个模型不在目标表上做梯度训练，而是把训练行当上下文，一次前向传播直接出预测，结果 14 场全胜。优势在数据量到 32000 行时依然成立。推理延迟每行 0.6 到 6 秒，表越宽越慢。另外，引用最多的那个 TabPFN ...
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageTabPFN with zero training beats tuned XGBoost on 14 out of 14 tables
The author tested TabPFN and TabICL against tuned XGBoost on 14 datasets from the Grinsztajn benchmark. Both models skip gradient training on the target table, using training rows as context in a single forward pass, and won all 14 matchups. The advantage holds up to 32,000 rows. Inference latency ranges from 0.6 to 6 seconds per row, with wider tables costing more. The post also notes that the most-cited TabPFN version now requires an account to download. The XGBoost tuning budget and hyperparameter search space are not detailed in the article.
Heat over time
Not enough continuous observations to draw a trend yet.